gemma-4-E2B-netsec-expert-GGUF

⚙️ Recommended runtime settings — gemma-native sampling temperature 1.0, top_k 64, top_p 0.95, min_p 0.01 (the min_p 0.01 floor prevents the reasoning-loop empty-answer issue), context length ≥ 16k (32k recommended), and a generous max_tokens when running with thinking on. The Gemma-4 thinking path needs --jinja.

gemma-4-E2B-netsec-expert (GGUF)

A ~2B edge model (gemma-4-E2B-it) with a Security & Networking FactBank baked into its chat-template. It answers correctly about post-cutoff / breaking-change APIs in 7 security & networking libraries — not by fine-tuning, but by carrying a searchable bank of landmine facts that fires inside llama.cpp at inference time. Weights untouched; no external RAG. Small enough to run on a laptop.

This is the most striking result of the set: a 2B model gains the most from the bank — it knows the least, so the bank fills the most gaps.

🔗 Full project — all experts, methodology, per-question transcripts, and benchmarks: github.com/mhndayesh/experts-models

What it fixes

114 curated landmine facts (post-cutoff · reverses-a-trained-habit · silent-failure) across: cryptography · OpenSSL 3 · paramiko 3 · urllib3 2 · volatility3 · yara-x · eBPF (BCC→libbpf).

Libraries in the bank (7) — 114 facts total

Where the facts come from (mined sources)

Each library's facts were extracted from its migration guide / changelog (source targeting is the whole game — a migration guide, not release-note noise), then quote-verified against the source line:

• cryptography changelog
• OpenSSL 3 migration guide
• eBPF BCC→libbpf migration guide
• paramiko changelog
• urllib3 v2 migration guide
• volatility3 migration guide
• yara-x differences doc

Full provenance (the mined source docs themselves) lives in the repo under v2/extractor/experts/security-networking/sources/.

Results (this model — hand-verified)

Same 48 landmine questions, base vs. this baked model, identical prompts (the bank injects in-engine):

+27 points — the biggest lift of the 2B/12B/26B curve (e2b 39.6→66.7% · 12B 56.2→81.2% · 26B 77.1→93.8%): raw lift shrinks with model size, so the edge model benefits most. Full methodology, transcripts, caveats: github.com/mhndayesh/experts-models → v2/extractor/experts/security-networking/.

How to run

The bank lives in the chat-template, so retrieval needs it applied — run on llama.cpp:

llama-server -m gemma-4-E2B-netsec-expert-Q4_K_M.gguf --jinja --port 8080 --ctx-size 8192

Query normally (same prompt as the base; the bank fires automatically for covered topics). Sampling — Gemma-native: temperature 1.0, top_k 64, top_p 0.95, min_p 0.01 (the min_p floor prevents reasoning-loop empty answers). For best accuracy send chat_template_kwargs={"enable_thinking": true} with an authority system prompt (tell the model the looked-up facts are verified and supersede its training) — a reasoning model otherwise reverts an injected fact to its trained prior. Don't quantize the KV cache on a model this small.

Limitations

• Scoped to the 7 covered libraries; outside them it's the base model.
• Supplies knowledge, not reasoning — a 2B still fails some multi-step transforms even with the right fact.
• Retrieval gate is token-based, with aliases. This bake includes the gate-alias fix — a natural or old name (e.g. "Volatility 3", RSA_new) also opens the tab; a wholly unrelated phrasing may still miss.
• Hand-scored landmine tests, not a general coding benchmark.

Also in this project — GitChameleon 2.0 vs. the frontier

The same FactBank approach on a different, code-execution benchmark: a local 12B + bank next to the published GitChameleon 2.0 leaderboard.

⚠️ Not apples-to-apples — the container harness was NOT run. Frontier = the official 328-problem run in pinned Docker containers (arXiv 2507.12367). FactBank rows = a local, non-Docker run over the 249 problems that built (hand-verified, some 3.7→3.9 remapped, no RAG). Base-vs-baked is internally fair; the frontier column is a different measurement — "what neighborhood," not a ranking. Details: LEADERBOARD-COMPARISON.md.

Papers

The write-ups behind this project (PDFs, rendered on GitHub): Research Report · Idea · Technical · Verdict · Evidence Ledger

Provenance & license

• Base: lmstudio-community/gemma-4-E2B-it-GGUF (Q4_K_M). This model = that GGUF with tokenizer.chat_template rewritten to embed an inverted-index retriever + the bank (factbank.version 0.4.0). e2b ships a different gemma-4 template revision than 12B/26B, so it required its own re-anchoring (see the repo).
• License: Google Gemma Terms of Use (license: gemma) — a gemma-4 derivative. The fact bank is from the FactBank project (repo LICENSE); mined sources keep their own licenses.
• Source & method: github.com/mhndayesh/experts-models.